Gemini 3.1 Flash Live ships with lower latency and smarter function calling

🧠 LAUNCH

Gemini 3.1 Flash Live ships with lower latency and smarter function calling.

Google's latest real-time audio model cuts response latency and adds reliable function calling mid-conversation — meaning your voice AI can now look things up, trigger actions, and stay coherent while doing it. The upgrade lands across Google products and AI Studio, making it the most accessible real-time voice API on the market right now. If you're building anything voice-first, this is your new baseline. (1,343 likes | 144 RTs) Read more →

Mistral drops Voxtral TTS: open-weight text-to-speech you can actually run. Mistral enters the TTS race with a 4B-parameter model that delivers realistic, expressive speech synthesis — and you can download the weights right now. Between quality that rivals closed APIs and a permissive license, this is the open-weight TTS moment the community has been waiting for. (2,694 likes | 108 downloads) Read more →

Cohere Transcribe claims SOTA for open-source speech recognition. Cohere Labs releases a new ASR model that benchmarks above Whisper on standard datasets. Combined with Voxtral TTS above, today marks the first time both ends of the speech pipeline — transcription and synthesis — have serious open-weight contenders shipping on the same day. (1,671 likes | 13 downloads) Read more →

Google Translate goes ambient: real-time translation through your headphones. Live translation now works through headphones on iOS and is expanding to more countries on Android. AI-powered translation is becoming invisible infrastructure — you don't open an app, you just hear your language. (Google Translate on iOS) Read more →

🔧 TOOL

Claude Code goes cloud-native: auto-fixes CI failures while you sleep.

Web and mobile Claude Code sessions can now follow your PRs and automatically fix CI failures without you sitting at a terminal. This is the async agent loop fully realized — push code, close your laptop, wake up to green checks. If you haven't tried cloud-based Claude Code sessions yet, this is the feature that makes them worth it. (3,769 likes | 283 RTs) Read more → — For a deeper look at why Claude Code is evolving beyond a traditional coding tool, see our analysis: Claude Code Is Not a Coding Tool.

OpenAI doubles Codex rate limits across all tiers this week. Every ChatGPT subscription tier gets 2x Codex CLI rate limits for the next week. If you've been hitting throttle walls while building, this is your window to push through. (991 likes | 33 RTs) Read more →

How Cursor built fast regex search for AI coding agents. Cursor's engineering team publishes their approach to indexing text for sub-second regex search — solving the bottleneck that slows every code-aware agent: finding the right code fast enough. Directly applicable if you're building any tool that needs to search codebases at agent speed. (24 likes | 5 RTs) Read more → — For context on how Cursor's approach compares to Claude Code's search strategy, see: Claude Code vs Cursor.

🔬 RESEARCH

Meta's TRIBE v2 creates digital twins of your brain's response to sight and sound.

Trained on 500+ hours of fMRI data from 700+ participants, TRIBE v2 generates zero-shot neural activity predictions for subjects it has never seen. That's not incremental — it means Meta has a foundation model that generalizes across human brains the way GPT generalizes across text. The implications for brain-computer interfaces, accessibility, and understanding cognition are enormous. (6,396 likes | 842 RTs) Read more →

Uni-1: a single model that both understands and generates. A team of ~15 researchers built Uni-1 to unify comprehension and generation in one architecture — no separate encoder and decoder models. The trend toward general-purpose multimodal models continues to accelerate; expect fewer specialized pipelines, more single-model stacks. (530 likes | 64 RTs) Read more →

DeepMind maps how AI can be weaponized for emotional manipulation. Google DeepMind publishes research cataloging AI misuse vectors across finance, health, and personal relationships — and proposes concrete safety measures. As voice AI gets more persuasive (see: every LAUNCH item above), this framework becomes required reading for anyone shipping user-facing AI. (200 likes | 31 RTs) Read more →

📝 TECHNIQUE

From zero to production RAG: an honest post-mortem of what actually breaks. A practitioner walks through building a production RAG system end-to-end — the non-obvious failure modes, the retrieval quality cliffs, and the evaluation gaps nobody warns you about. This is the kind of hard-won wisdom that saves you weeks of debugging. If you're past the demo stage and hitting real-world RAG issues, start here. (274 likes | 84 RTs) Read more → — For more on RAG architecture patterns and common pitfalls, see: RAG.

Voice-controlled app building with Gemini Flash Live in AI Studio. Google demos talking through an app build in real time — you brainstorm out loud and the model keeps pace, generating code as you think. It's rough around the edges, but the low-latency voice loop hints at a future where the keyboard is optional for prototyping. (289 likes | 48 RTs) Read more →

💡 INSIGHT

Karpathy: AI writes the code, but the infrastructure maze remains unsolved.

Andrej Karpathy reflects on building menugen a year ago and names the real wall every vibe-coder hits: services, payments, auth, databases, domains, DNS — the full DevOps stack that no AI can navigate yet. His conclusion: the next unlock isn't smarter code generation, it's agents that handle deployment end-to-end. He's right, and whoever solves this owns the next wave of AI-native development. (2,834 likes | 212 RTs) Read more →

Minute-by-minute: how one developer caught the LiteLLM supply chain attack. A detailed incident response timeline from someone who detected the LiteLLM malware in real time — from first anomaly to full containment. This isn't theoretical security advice; it's a concrete playbook for what to do when an AI dependency gets compromised. Print it out and tape it next to your monitor. (269 likes | 120 RTs) Read more →

Anthropic throttles Claude sessions during peak hours to manage demand. 5-hour session limits for free, Pro, and Max users now compress during weekday peak hours (5am–11am PT). Weekly limits stay the same. Translation: schedule your heavy Claude sessions for evenings or weekends, or upgrade to Team/Enterprise for unthrottled access. (3,553 likes | 236 RTs) Read more →

Federal court grants Anthropic injunction against U.S. Department of War. A federal judge issued a preliminary injunction in Anthropic v. U.S. Department of War — a significant legal precedent for AI companies pushing back on government mandates. The ruling's reasoning on AI autonomy and corporate ethics obligations will shape policy for years. Read the actual document, not the summaries. (28 likes | 2 RTs) Read more →

🏗️ BUILD

AI-assisted JSONata rewrite: one day of work, $500K/year saved. Reco.ai used AI to rewrite their JSONata dependency in a single day, eliminating a half-million-dollar annual cost. The real lesson isn't the dollar figure — it's that AI-assisted rewrites have crossed the threshold where tackling gnarly legacy dependencies is now a one-day project, not a one-quarter project. Identify your own rewrite candidates. (55 likes | 49 RTs) Read more →

🎓 MODEL LITERACY

Zero-Shot Generalization: When TRIBE v2 predicts brain activity for a subject it has never been trained on, that's zero-shot generalization — the ability to perform well on entirely new inputs without any fine-tuning or additional training data. Today's open-weight audio models do the same thing: Voxtral TTS handles voices and accents it never saw during training, and Cohere Transcribe recognizes speech patterns from unseen speakers. The mechanism is similar across domains — during training, the model learns general patterns (how brains respond to stimuli, how speech sounds map to text) rather than memorizing specific examples. Understanding zero-shot generalization is the key to evaluating any model's real-world utility: benchmarks test what the model has seen, but your production data is what it hasn't.

⚡ QUICK LINKS

Google Search Live: Expands to all AI Mode languages and locations globally. Link
$7/month AI agent: A working agent on a cheap VPS with IRC as its transport layer — beautifully scrappy. Link
Mollick on AI ambition: "If companies aren't failing with AI, they're not being ambitious enough." (405 likes | 45 RTs) Link
Latent Space: Breaks down Anthropic's biggest launch week ever — computer use, auto mode, scheduling. Link
The AI dev inner loop: The emerging workflow patterns that separate productive agent users from frustrated ones. (85 likes | 9 RTs) Link
Manyika × LL COOL J: Google's SVP of Technology and Society talks AI and creativity with a hip-hop legend. Link

🎯 PICK OF THE DAY

TRIBE v2's zero-shot neural predictions aren't just a neuroscience flex — they're proof that foundation model scaling laws apply to biological data. Meta trained on 500+ hours of fMRI from 700+ people and got a model that predicts brain activity for subjects it has never seen. Read that again: the same scaling playbook that gave us GPT and Claude — more data, bigger model, emergent generalization — works on the human brain. The team that cracks brain-to-model alignment first doesn't just advance neuroscience; they own the next input modality after voice. Today's open-weight audio launches (Voxtral TTS, Cohere Transcribe) already show zero-shot generalization across unseen voices and accents. TRIBE v2 extends that same principle to neural signals. We're watching the convergence of foundation model techniques across text, audio, and now brain data — and the implications for brain-computer interfaces, accessibility, and eventually direct neural input are staggering. This is a slow-burn story that will look obvious in hindsight. (6,396 likes | 842 RTs) Read more →

Until next time ✌️